# TareKV

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Haiming Ni](https://cerebralvalley.ai/u/nniiovo)
- **GitHub:** https://github.com/small-amumu/kcache-tools
- **Demo video:** https://youtu.be/f0PbMT6812I?si=JCuCdwY5y_ktWGdL
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5127

TareKV helps developers shrink AI memory without silently breaking answers. It checks a compressed cache, tries a shared-pattern repair when needed, and checks again. If it still fails, it falls back to higher precision.

Our open-source tools support Qwen and Meta Llama in llama.cpp. The repository reports Qwen2.5-7B prediction agreement recovering from 6.5% to 95.0% with q4_0 repair. Our three-minute visual demo explains the mechanism and replays these reported results.

## More from Meta Global AI Developer Hackathon

- [Arupa AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5124)
- [Vishisht Magan](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5125)
- [Ethan Gritli](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5126)
- [Adam Je](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5128)
- [SyncDigit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5129)
- [OurTake](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5130)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5127. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt